Certification of machine learning applications in the context of trustworthy AI with reference to the standardisation of AI systems

Abstract

Artificial intelligence (AI) and its subset machine learning (ML), which focuses on learning tasks from data, are some of the most disruptive emergent technologies. AI standards inform organisations how to develop and manage their AI systems and are emerging to satisfy the increasing demand from industry and society for the safe adoption of AI and ML technologies. However, AI systems must be trustworthy in the sense that they can be relied upon to make responsible decisions. Consequently, trustworthy AI is a collection of principles that encourages responsible development, use and deployment of AI systems, and can be viewed as a framework for managing risk in AI systems. The National Physical Laboratory (NPL) is one of the four institutions responsible for delivering the UK’s national quality infrastructure (NQI), in which standards and certification play key roles. In this context we review research in NPL on trustworthy AI, emphasising the importance of uncertainty quantification (UQ) in enhancing the transparency and trust in results output from AI systems. We review the landscape of AI standards and certification and emphasise their role in the context of trustworthy AI. Third-party certification is a key service in building trust in AI and ML systems and supporting their operationalisation. We argue that certification should assess conformity to AI standards and characteristics of trustworthy AI, and, in addition, should be able to carry out conformity testing and evaluation of the components of an AI system. As a case study we look at ChatGPT, a large AI system which is attracting a lot of attention, and investigate its potential conformity to the principles of trustworthy AI.